Towards the Detection of Various and Extremely Small Wires in Cables

Understanding the behavior of wires in cables is essential for assessing the quality of cable products. Object detection or segmentation technology is typically employed to identify wires depicted in Computed Tomography (CT) images. Despite advancements in machine learning-based object detection, wire detection remains a research challenge owing to cables’ numerous and varied shapes and sizes. Each cable type requires a different annotated dataset, creating a significant annotation burden. This paper proposes an integration model combining self-supervised learning (SSL) and multiscale segmentation (MSS) for wire detection. The SSL approach generates feature maps of wires without needing annotated datasets, while the MSS handles images with different resolutions. By leveraging both methods, we successfully detected wires in five different types of cables using annotated images from only one cable type, reducing the number of annotated images needed without compromising detection accuracy.